Four curated sequences

Start with the first-quarter rollout: agree on data and permission rules, define a pilot, establish checks, and review the results.

Use the one-page policy to agree on ownership, data access, verification, and escalation before the pilot.

The patterns inside the repo that make the practice durable: instruction hierarchy, hooks, shared memory, and workflows as version-controlled processes.

The loop itself — draft, verify, revise — applied specifically to agents, with the verification strategies that separate generators from agents.

Keep time for independent diagnosis and review, especially when an assistant handles unfamiliar code.

The overview of the audit itself — the four tracks, what each one delivers, and why you do them in parallel rather than in series.

Build a workload-level cost baseline, identify gaps in aggregate billing and utilization views, and record the assumptions.

The reliability track. A taxonomy of what breaks on GPUs in production — from thermal throttling to VRAM fragmentation — and how to catch each early.

The SLO track. How to define meaningful latency, error, and saturation targets for inference, and what to do when the error budget is spent.

Move from the audit to operating procedures: pinned configurations, repeatable deployments, and documented recovery.

Start with the national context. What changed between Meaningful Use and TEFCA, and what still blocks plug-and-play apps despite a decade of standards work.

The operational playbook: contracts as assets, observability first-class, executable docs, and support workflows that compound.

A specific failure mode — an "enhanced best match" trap where filter and mode parameters collide — and the defensive pattern that fixes it on both sides.

The reference implementation. How Source Profiles, a three-phase parsing pipeline, and a workflow DSL turn messy legacy formats into semantic events without a one-off codebase per feed.

Start with the shape of a reliable agent: an explicit loop, bounded tools, verification against concrete criteria, and visible operator state.

The repo as control system. Hooks, memory, tasks, and workflows that give agents durable context and enforce the invariants you care about.

Maintain engineering judgment through independent checks, explanation, and occasional work without the assistant.

The policy that makes the practice auditable. Ten lines the team converges on, with a quarterly review rhythm.